Tracking the heart motion during radiation treatment of cancer patients can provide important information for designing strategies to reduce radiation-induced heart toxicity. Recently, in-treatment cine MRI images are used for guiding radiation therapy. However, dynamic changes of heart shape and limited-contrast of cine MRI images make automatic heart motion tracking a very challenging task. This paper proposes a deep generative shape model-driven level set method to address these challenges and automatically track heart motion on 2D cine MRI images. First, we use a three-layered Deep Boltzmann Machine (DBM) to train a heart shape model that can characterize both global and local heart shape variations. Second, the shape priors inferred from the trained heart shape model are incorporated into the distance regularized level set evolution-based segmentation method to guide frame-by-frame heart segmentation on cine MRI images. We demonstrate the superior performance of the proposed method on cine MRI image sequences acquired from seven volunteers and also compare it with four other methods.